Choursi et al. (2026) Ensemble Machine Learning for Meteorological Drought Assessment and Forecasting with Satellite and Climate Data (Urmia Lake Basin, Iran)

Identification

• Journal: Water Cycle
• Year: 2026
• Date: 2026-02-01
• Authors: Sima Kazempour Choursi, Mahdi Erfanian, Hirad Abghari, Mirhassan Miryaghoubzadeh, Khadijeh Javan
• DOI: 10.1016/j.watcyc.2026.02.002

Research Groups

• Department of Range and Watershed Management, Urmia University, Urmia, Iran
• Department of Geography, Urmia University, Urmia, Iran

Short Summary

This study developed and evaluated ensemble machine learning models, particularly Extremely Randomized Trees (ERT), for meteorological drought assessment and forecasting in the Urmia Lake basin, demonstrating superior accuracy and the shifting influence of local vs. teleconnection drivers across different timescales. The ERT model consistently outperformed other algorithms, providing reliable 3–6 month drought forecasts with high accuracy.

Objective

• To assess and forecast meteorological drought in the Urmia Lake basin using ensemble machine learning models, integrating satellite data, meteorological observations, and climate teleconnection indices across various timescales (3-, 6-, 9-, and 12-month intervals).

Study Configuration

• Spatial Scale: Urmia Lake basin (approximately 52,000 km²), with all spatial data resampled to a common 5 km resolution.
• Temporal Scale: Study period from 2001 to 2019 (19 years), with monthly data and drought indices computed for 3-, 6-, 9-, and 12-month accumulation periods, and forecasts generated with 3- and 6-month lead times.

Methodology and Data

• Models used: Decision Tree (DT), Random Forest (RF), Extremely Randomized Trees (ERT).
• Data sources:
  • Satellite:
    • Tropical Rainfall Measuring Mission (TRMM) 3B43 Version 7 (precipitation, 0.25°).
    • GPM Integrated Multi-satellite Retrievals for GPM (IMERG) (precipitation, 0.1°).
    • Global Satellite Mapping of Precipitation (GSMaP) (precipitation, 0.1°).
    • MODIS Collection 6 Level 3 products: MOD11C3 V061 (day/night land surface temperature, 0.05°), MOD13C2 V061 (Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), 0.05°), MOD16A2GF V061 (potential evapotranspiration (PET), 500 m).
    • ASTER Global Digital Elevation Model (GDEM) Version 3 (elevation, 30 m).
  • Observation/Reanalysis:
    • Monthly precipitation records from 12 synoptic meteorological stations (Iran Meteorological Organization).
    • Global Precipitation Climatology Centre (GPCC) gridded precipitation product (V2022, 0.25°).
  • Teleconnection Patterns:
    • Southern Oscillation Index (SOI).
    • Multivariate El Niño–Southern Oscillation Index (MEI).
    • North Atlantic Oscillation (NAO).
    • Atlantic Multidecadal Oscillation (AMO).
  • Drought Indices (calculated): Standardized Precipitation Index (SPI), Standardized Precipitation Evapotranspiration Index (SPEI).
  • Ancillary: Categorical Month variable.
  • Analysis Tools: SHapley Additive exPlanations (SHAP), Wavelet Coherence Analysis, Monte Carlo Simulation.

Main Results

• The Extremely Randomized Trees (ERT) model consistently demonstrated superior performance over Decision Tree (DT) and Random Forest (RF) models in both calibration and validation for SPI and SPEI estimation across all timescales (3, 6, 9, 12 months), achieving higher R² values (e.g., up to 0.987 for SPEI3) and lower RMSE/MAE.
• ERT achieved the highest spatial agreement (Kappa coefficients predominantly exceeding 0.8) and classification accuracy (Overall Accuracy consistently above 0.94) for drought severity.
• Input variable importance analysis (SHAP and relative importance) revealed a clear shift in dominant predictive drivers:
  • Short-term droughts (3-6 months) were primarily controlled by local hydrological variables such as TRMM precipitation, daytime land surface temperature (LSTday), and Normalized Difference Vegetation Index (NDVI).
  • Longer-term droughts (9-12 months) were more strongly influenced by large-scale climate teleconnection indices (MEI, SOI, AMO, NAO).
• Wavelet coherence analysis confirmed these multiscale controls, showing strong and persistent coherence of TRMM precipitation, LSTday, and NDVI at short-to-seasonal scales (8-32 months), while teleconnections emerged as significant modulators at mid-to-longer scales (16-32 months).
• The ERT model successfully generated spatially explicit drought severity maps and forecasts with 3- and 6-month lead times, showing strong agreement with observed data (R² values exceeding 0.95 for forecasts).
• Monte Carlo simulation confirmed the ERT model's high classification stability and noise resilience, with the probability of consistent drought classification across simulation runs exceeding 0.85 at most stations and months.

Contributions

• First systematic application of Decision Tree (DT), Random Forest (RF), and Extremely Randomized Trees (ERT) for spatial Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI) forecasting in the Urmia Lake basin.
• Integration of multiscale SHapley Additive exPlanations (SHAP) attribution and wavelet coherence analysis to elucidate the shift from local to teleconnected drought drivers across different timescales.
• Rigorous categorical and uncertainty validation (Cohen’s Kappa, Monte Carlo resampling) to support operational early warning and water-allocation decisions.
• Development of a robust and transferable framework for basin-scale drought assessment and early-warning systems, particularly applicable to semi-arid regions.
• Contributes to UN Sustainable Development Goal 6 (Clean Water and Sanitation) and SDG 13 (Climate Action) by advancing methods for proactive water-resource management and enhancing resilience to climate-driven drought.

Funding

The authors affirm that no funding, grants, or external financial support were received for the preparation of this manuscript.

Citation

@article{Choursi2026Ensemble,
  author = {Choursi, Sima Kazempour and Erfanian, Mahdi and Abghari, Hirad and Miryaghoubzadeh, Mirhassan and Javan, Khadijeh},
  title = {Ensemble Machine Learning for Meteorological Drought Assessment and Forecasting with Satellite and Climate Data (Urmia Lake Basin, Iran)},
  journal = {Water Cycle},
  year = {2026},
  doi = {10.1016/j.watcyc.2026.02.002},
  url = {https://doi.org/10.1016/j.watcyc.2026.02.002}
}